Parallel Generative Topographic Mapping: An Efficient Approach for Big Data Handling.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Generative Topographic Mapping (GTM) is effective for analyzing large chemical datasets.
• Constructing GTM requires a representative Frame Set (FS) of compounds.
• Training GTM becomes computationally infeasible with very large FS sizes.

Purpose of the Study:

• To develop a Parallel GTM algorithm to overcome computational limitations with large FS.
• To assess the efficiency of the Parallel GTM algorithm across various FS sizes.
• To determine the optimal FS size for representing large chemical spaces.

Main Methods:

• A Parallel GTM algorithm was proposed, merging intermediate manifolds from molecular subsets.
• 80 GTMs were trained on FSs ranging from 10 to 1.8 million compounds from the ChEMBL database.
• Classification models for up to 712 biological activities were built to evaluate GTM performance.

Main Results:

• The Parallel GTM procedure successfully covered the full spectrum of FS sizes.
• A FS of 5000 randomly selected compounds was proven sufficient to represent 1.8 million ChEMBL molecules.
• Increasing FS size beyond 5000 compounds showed no significant improvement in predictive accuracy for biological activities.

Conclusions:

• A small, representative FS is sufficient for accurate predictive modeling of large chemical libraries.
• Parallel GTM enables the mapping of vast chemical spaces, including billions of compounds.
• The findings challenge previous assumptions about FS size requirements in GTM.